I’m all for championing excellence over mediocrity, but it is a mistake to imply that academic achievement is the only measure of it. Jocks and prom queens aren’t the antithesis of merit and striving, and in some ways, their arenas are more “real” than academic curriculums. Fight passivity and low standards!

AI may well be able to write massive software projects “from scratch”, but that almost certainly isn’t the fastest / highest performance / highest security / most token efficient way to do things. Good building blocks matter. In addition to RL’ing the agents on final task loss, we should try to get them to regularize the building blocks as well. What would the standard library look like that makes the agents most effective? It is likely at the limit that it turns out to be some alien organization of code, but given the existing human-bias in training, I bet that in the near term it would be lovely code that old programmers would appreciate. I suspect the same could be said for mathematics, but that is out of my lane.

I see how some people can watch flashy robot demos on the timeline and think with a shiver that the robot apocalypse is around the corner, but reality check: There is no robot in the world that can change the batteries in a jack-o-lantern. If you have a robot friendly door handle it might be able to walk outside, pick it up, and carry it back inside. The robot could open the drawer with your batteries in it, but then things get tough. Opening the package of AA batteries? Extracting a precision screwdriver from its case? Unscrewing four tiny screws that require a hard press and twist? Popping batteries out of the case? Yes, you could build and train a specialist robot to do that today, but there is a very long way to go to replace the full range of human capabilities. Near term doomer scenarios should stop at “secretly controlling humanity” instead of “wiping humanity out” 😀

That looks really neat! The automatic analysis and use of multiple data stacks and resizable fields takes it from "lots of work can be done this way" to "basically everything can be". The relative references and built in binary structure validation also enable direct memory mapping of data structures and immediate use of network data. "Parsing" file and network data is often 100x slower than it needs to be.

I had a similar reaction. Everyone should applaud performance improvements, but gauge against the “speed of light”, not just a sloppy competitor. If you are measuring in seconds, you may have already anchored yourself incorrectly. Structure the UI so you can put locally cached media up just a few vblanks after the process launches. Structure the server so it can return everything necessary for the dynamic UI in one single query that was kicked off in parallel with the Gfx API initialization, branching to a modern transport protocol for large media files. You should be able to double click on the launcher and have it responsive before your muscle memory can slam the pointer to a location to click on.

Waymo and Zoox first impressions I use self driving on my Tesla all the time, but I finally got around to trying the commercial autonomous ride hailing services this week in Las Vegas. The biggest issue by far is the very limited pickup and dropoff points, and only having a couple minutes to get to them resulted in a bit of a rush. If you are going to have a small-integer number of locations, you should make it super-obvious where those locations are. Waymo included a picture, but it was a rather undistinguished parking garage image and it wasn’t clear which floor it was on. Being able to switch the map to an aerial photography view might be helpful in some cases. Pan-able 360 photos of the target points might also be helpful. The first Zoox ride was flawless, but we had a lengthy walk to get to our actual destination. The first Waymo ride was a little rougher. It felt more indecisive with steering, like Teslas did a couple years ago. This was still early-access in Vegas, so it is probably better in the active commercial markets. The second Zoox ride had a small incident – there was another Zoox ahead of us at a stoplight (and a Waymo and another Zoox next to us!) that repeatedly tried to back up into our Zoox, to the point that ours actually honked at the one ahead of it a couple times. Eventually, after multiple light cycles, ours started to tentatively try to get around it, which left it partway into the next lane and distressingly close to a fast moving semi. The one ahead finally moved, and the ride continued normally. Overall, the Zoox experience felt a bit more charming than the Waymo, with several nice little touches, but they both get the job done. I still need to try a Cybercab.

Whenever I see a decorative mirror ball, I think about the old-school OpenGL GL_SPHERE_MAP texgen option, and light probes for realistic lighting of synthetic renderings for merging with photos. Long before cube maps were an option, sphere maps allowed a form of shiny surface rendering, but only from a limited viewpoint — you couldn’t walk all the way around a shiny object. For light probes, you could take multiple photos with different exposure times to capture a true many-orders-of-magnitude high dynamic range composite image for realistic image based lighting of everything from direct sunlight to subtle shadowed region indirect lighting.

After playing some pinball on our Polycade cabinet, Trista and I revisited VR pinball on our Quest 3, and we both thought it was a meaningfully better and more fun experience in VR. There is an interesting wrinkle: @PinballFX used an efficient custom engine for a long time, but they switched over to Unreal Engine for their current games to more easily support the wide range of platforms they deploy on. Pinball FX uses Unreal, but they left the older version available as PinballFX Classic, which makes for a nice comparison. After downloading the 40x larger new application, you find that it has tons of customization features and environmental effects, but most tables drop frames even on Quest 3. The Classic version is locked at an unfortunately low resolution, but it feels more solid when actually playing pinball. Quest Game Optimizer can adjust resolution and frame rate if you are willing to go through the Side Quest dance, making a fairly dramatic impact on the visuals. I just gave Meta a nudge about maybe allowing QGO into the official store, but ideally the developers would add a few lines of cold to the old application to default to different settings for the different Quest headsets. Touching an old application is a bigger ask for a developer than most users appreciate, though. Both the old and new versions are kind of a mess with preventable visual aliasing. Here is a post I wrote ten years ago about aliasing for GearVR developers, long before Quest existed: sRGB framebuffers and sRGB textures with mip maps, MSAA, go light on specular shading effects. < Dr Manhattan it’s 1985 meme >

I recently had a conversation with a NASA JPL engineer that was bemoaning the fact that the internal “Planetary Protection” department at NASA threw up substantial roadblocks for Mars missions in the name of protecting the pristine Martian environment. They will surely be apoplectic about this. I kind of think we should just get it over with and empty a septic tank on mars so nobody can imagine it needs dedicated protection. Project Vandalize Mars.

Very few will care about the actual reinforcement learning details here, but watching me grope around for understanding in the chat log might be of broader interest: I have often noticed that our estimated Q values are higher than the observed returns, sometimes substantially so, and that can’t be good for performance. For policy decisions, only the relative values matter, but an offset negatively impacts bootstrapping, so I was excited to see this paper on Relative Value Learning: After struggling a bit with “the Banach space of bounded antisymmetric pairwise functions“, I realized all that is essentially going on is subtracting two value functions. Given this framing, Chat was able to simplify the machinery into a particularly elegant form: Subtracting the mean TD error from the individual sample TD errors gives all the benefits of relative value learning. Summarized as "relative Bellman regression is TD-error centering". Unfortunately, while you can create examples where this should be very valuable (someone should write a proper paper on it!), it didn’t actually improve performance on my tasks. However, I think this has usefully narrowed down what is actually happening. Bellman iteration naturally corrects towards a correct absolute value, but only when the bootstrap values are also training targets. With Q-learning, you are often / mostly bootstrapping from a max-action that was not actually taken, so it never sees any downward Bellman pressure. This is consistent with another result I have noted: learning state value functions offline from a frozen policy without actions doesn’t seem to suffer from any value overestimation.

I recently went through a translation of Musashi’s Book of Five Rings. The introduction chronicled the evolution of swordsmanship and martial arts in general from pragmatic battlefield necessities to sports, historical curiosities, and hobbies. It was interesting to hear how in the post-WW2 occupation when martial arts were banned, Judo was represented as equivalent to western wrestling, and Kendo later to western fencing. The poignant undercurrent for me is that I can see many programming skills following that path. There are probably dozens of people reading this that remember hand assembling opcodes to hex and still have some magic numbers burned into their memory, but even the small group of people still programming in assembly today (hey, @FFmpeg !) don’t work at that primitive level now. AI is making many other programming skills much less critical. We aren’t there yet, but carefully writing code completely by hand is moving from a -jitsu to a -do. Code-do? Codo? That’s ok! The retro computing scene is delightful, full of people building and exercising old skills for the love of it. But don’t be the out of touch Kung Fu master, heir to lifetimes of tradition, that gets mauled by an amateur MMA fighter. Musashi would probably have been pretty enthusiastic about assault rifles.

If the target market is real time robotics, the Jetson Thor system, with 128 GB of memory, but only 273 GB/s of bandwidth, seems over-provisioned with expensive memory. You want the model evaluating at tens of fps, so it can't use more than 10 GB of weights at most. More memory could be used if the model is a wide mixture of experts, or a large strategic planning model is operating in small time slices over a longer period of time, and more memory always makes development life easier, but for cost optimized systems, you should be able to get by with much less.

Now that serious industrialization of space in the name of data centers is within the Overton window, it feels like time to bring up Rotovators again. Even if Starship completely meets all design goals with low overhead orbital reusability, it isn’t the end-all of cost effective access to space. Momentum exchange tethers could substantially reduce propellant consumption and operational challenges. The idea is that you have a tapered cable a few hundred km long attached to a substantial dead weight, and you set it spinning like one spoke of a wheel as it orbits earth. A rocket could fly up to space at much less than orbital velocity and grab onto the end of the rotating tether, then get pulled up to a faster speed and fling the payload off into a stable orbit. Falcon precision landings (with suicide burns!), and now SuperHeavy catches demonstrate that big rockets can precisely dock with infrastructure, and doing it in space is generally easier than doing it on the ground. The big win is in electrodynamic tethers, where you can theoretically use (lots and lots of) solar panels to raise the tether orbit after it has pulled itself lower while picking up a payload. That would be a game changer, but it hasn’t been unequivocally demonstrated to be practical. Even without that, just using chemical propulsion, it could still give significant efficiencies. Only the mass of the payload that is actually flung off the top of the tether needs to have the velocity made up by engine burns. The rest of the mass of the rocket “gives it back” when it drops back off at the bottom several rotations later over the launch site. Because the pickup and dropoff happen at much less than orbital velocity, there is no terrible reentry heating challenge. The vehicle gets simpler and more robust. So, you could rebalance a two stage rocket to be more propellant efficient, or you could make a reusable single stage to tether rocket. A SSTT might still use more propellant per kg of payload than Starship, but the operational advantages would be large. You lose lots of flexibility with tether systems compared to conventional rockets, but it would be like putting deep water harbors in key places.

I still see the Scratch “visual programming” environment positioned as an introductory path to programming for kids, but unlike game modding, I have never heard a “success story”, where someone credits their early experience with Scratch as key to a fulfilling career. Anyone?

Trista gave me personalized AI generated comic books and bookmarks on my birthday this year, and they made me smile. I noted a couple places where the rendering glitched or the model lost a little bit of coherence in the story, but these have come a long way from the early on-demand kids books created with traditional coding practices.

I finished @MorlockP ‘s Red State Mars last week, and I thought the afterword offered a really interesting insight into the creative process behind the book. Travis is a tech guy (and state representative) that writes “Good Old Fashioned Science Fiction”, and I have enjoyed his work. If you are inclined that way, go ahead and read the book. If you are put off by the implied political bias, maybe read the afterword and see if the themes sound interesting. I wound up being briefly referenced in the context of zero-G combat, since Russ Blink and I did try a little bit of judo on a parabolic airplane flight long ago. Sixty seconds of experience puts us ahead of at least 99.999999% of the population, but I bet there are some Russian cosmonauts that have done some more serious and extended tussling, which would be fun to hear about in an interview sometime.

Whenever I look at a building wall with a random-seeming mix of two brick colors, I think about decoding it as a bit stream. Surely a university building somewhere has put something meaningful in the pattern.